Cancelable Multi-Branch Deep Learning Framework for Privacy-Preserving ECG Biometric Authentication

Recently, biometric authentication systems have become essential tools for secure and reliable identity verification. Electrocardiogram (ECG) biometrics are particularly effective because of the unique electrical activity patterns of the heart. However, storing and processing raw ECG data raises privacy and security concerns, requiring advanced methods to safeguard user data. This paper presents a novel cancelable ECG biometric authentication framework that integrates a Cancelable Dynamic Biometric Transformation (CDBT) with a multi-branch deep learning architecture to enhance both privacy and performance. The proposed model captures varied ECG feature representations via parallel convolutional branches and transforms them into non-invertible templates using a fixed-key CDBT method. We evaluate the framework on two publicly available ECG datasets (PTB and ECG-ID) achieving high authentication performance with accuracies of 96.25% and 96.43%, respectively, and Equal Error Rates (EER) of 3.55% and 3.6%. The system meets critical privacy-preserving criteria, including irreversibility, unlinkability, diversity, and revocability. In addition, the model demonstrates robustness against spoofing and transformation-based attacks. These results confirm the practicality and scalability of our approach in secure, real-world biometric applications.